Claude Fable 5: Anthropic opens its most powerful model to the public

Today Anthropic released Claude Fable 5 - and for the first time, its most capable tier of model is open to the public. Fable 5 is a Mythos-class model, a tier that sits above the Opus class in capability, and the same frontier generation Anthropic had, until now, kept out of general use because it was simply too capable to hand out freely.

The release threads a needle Anthropic has worked at all year: how do you ship a model this capable without also shipping its most dangerous uses? The answer here is not a watered-down model. It is a safety net.

A model held back - until now

Anthropic introduced Mythos in April and said straight away that it would not make the model broadly available. Mythos was unusually good at things like finding flaws in software, a skill that helps defenders and attackers alike. So instead of a public launch it went to a small, vetted group under a programme called Project Glasswing, built around cyber defenders and critical-infrastructure providers - AWS, Microsoft, Apple, CrowdStrike and others.

Claude Fable 5 is the public face of that same underlying model: frontier capability, wrapped in guardrails that make a wide release possible.

The idea: keep the power, fence off the danger

Here is the move. Rather than dialling down what the model can do, Anthropic kept Fable 5 at full strength and built safeguards around it that behave like a safety net. For the overwhelming majority of requests - coding, writing, research, analysis - you get the model's complete capability. But for a narrow set of high-risk topics - cybersecurity, biology and chemistry, and model distillation - Fable 5 does not answer directly. Those requests are handed to Claude Opus 4.8, Anthropic's previous flagship and an already-public model, which returns a safe response, and you are told when it happens.

In practice the net is rarely triggered. Anthropic says more than 95% of Fable sessions never hit a fallback at all - and when they do not, Fable 5 performs essentially the same as the unrestricted Mythos 5.

What Fable 5 can do

Anthropic says Fable 5 surpasses every model it has released to the public, and its headline strengths cluster around long, complex, hands-off work.

On coding, it is built for scale - large migrations and intricate implementations - and can run on its own for days inside an agent harness like Claude Code, planning across stages, handing subtasks to sub-agents, and writing its own tests. One launch example: Stripe used it to complete a migration across a 50-million-line Ruby codebase in a single day, work that would have taken a team more than two months.

On vision, Anthropic calls it the new state of the art - it can pull precise numbers out of detailed scientific figures and rebuild a web app's source code from screenshots.

And on everyday knowledge work, it beat Opus 4.8 on Anthropic's spreadsheet suite at every effort level while finishing runs 25 to 30 percent faster, using fewer turns. One early enterprise tester, the trading firm IMC, reported it aced their trading-analysis evaluations nearly across the board.

The bet is that you can put a frontier model into millions of hands and still hold its sharpest edges back.

How safe is the safety net?

A guardrail is only as good as its weakest point, so Anthropic put Fable 5 through both internal testing and outside scrutiny. It says an external bug bounty ran more than 1,000 hours against the safeguards without producing a universal jailbreak - though it is candid that the UK's AI Security Institute made progress toward one in a short window, so the honest framing is "so far", not "never". Using a Mythos-class model also comes with a condition: 30-day retention on all traffic, including for enterprises that previously had zero-retention agreements. Anthropic says that data is used only to defend against novel attacks and reduce false positives, and never to train future models.

Mythos 5, for a chosen few

Alongside Fable 5, Anthropic also released Claude Mythos 5 - the same model with the guardrails lifted - but kept it restricted. Project Glasswing partners running the preview can upgrade now, with the cyber safeguards removed; selected biology researchers are next, through a forthcoming trusted-access programme; and Anthropic says it will widen access from there.

How to try it

Claude Fable 5 is available today on the Pro, Max, Team, and seat-based Enterprise plans at no extra cost through June 22. On June 23 it moves to usage credits, to return as a standard plan feature once capacity allows. Developers can call it through the Claude API as claude-fable-5, and it is also on Amazon Bedrock, Google Cloud's Vertex AI, Microsoft Foundry, and GitHub Copilot. Pricing is $10 per million input tokens and $50 per million output tokens - double Opus 4.8 - with cached input reads discounted by 90%.

What this means if you are building

For teams building on Claude, the practical headline is simple: frontier-grade capability is now a normal API call, on the same clouds you already use. The safety net mostly stays out of the way, and when it does engage it does so predictably, on a known set of high-risk domains. It also sits at the opposite end of the market from an open model you run yourself: Gemma 4 12B on a laptop trades raw capability for control and privacy, while Fable 5 trades a little gating for the strongest model Anthropic has shipped publicly. Most real products will draw on both. If the guardrails hold, today is a preview of how frontier AI reaches everyone without the worst-case risks riding along.